acceptodds
Under review as a conference paper at ICLR 2027

Generating Heavy-Tailed Financial Returns via Elliptical Variational Auto-Encoders

Abstract

Financial risk management often involves learning the return tail distributions, which include both market co-movement and asset-specific extremes. Inspired by the elliptical approximate factor model in the financial literature, we develop an elliptical variational auto-encoder (E-VAE) based on its nonlinear extension, which decomposes financial returns into a nonlinear common factor component and idiosyncratic fluctuations; both follow heavy-tailed distributions with occasional jumps. The elliptical VAE admits a standard evidence lower bound with analytic Kullback–Leibler terms and reparameterized training. In financial applications, the fitted encoder provides posterior estimates of the common factors from historical observations, while the fitted decoder can generate synthetic return vectors for downstream risk assessment and portfolio planning tasks. By exploiting the low-dimensional factor structure, our model enables efficient sampling of high-dimensional return vectors. Experiments on synthetic data and the U.S. and Chinese stock markets demonstrate its ability to capture both the body and tails of return distributions, supporting its use in downstream financial decision-making.

Primary area:
generative models
Keywords:
Financial generative models, Heavy-tailed distributions, Nonlinear factor models, Variational autoencoders

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ARHsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a generative model, E-VAE, tailored to learning the tail distribution of financial returns. To this end, the method adopts Student's tt-distributions for both the decoder and the prior, while additionally incorporating jump distribution modeling. Two types of jump components are introduced: common and private jumps. These components are parameterized using Bernoulli distributions, resulting in mixture formulations for both the decoder and the prior. The paper empirically demonstrates that the proposed model achieves competitive performance in distribution fitting as well as in a downstream portfolio evaluation task.

Critical strengths
  1. The paper derives an analytically tractable training objective while jointly incorporating heavy-tailed distributions and explicit modeling of jump behavior.
Critical weaknesses
  1. Training the proposed model appears to require tuning a relatively large number of hyperparameters. Appendix E.3 lists nine hyperparameters, including mixture probabilities, relative scale parameters, and degrees of freedom. This may result in a substantial computational burden during hyperparameter selection and raises concerns regarding the practical efficiency and robustness of the proposed method.

  2. In the synthetic-data experiments, the comparison with the baseline methods is currently presented primarily through the visualizations in Figure 1. Quantitative evaluation would be necessary to more objectively assess the differences shown in this figure. For example, statistics based on the Kolmogorov-Smirnov test could be used to summarize the distributional discrepancies illustrated in the bottom panels of Figure 1.

  3. Theorem 2 is introduced in the middle of the experimental section without citation, but its role in the overall contribution is not sufficiently explained. In particular, it is unclear (1) why this theorem is theoretically meaningful for the proposed method and (2) how its result is connected to, or supported by, the subsequent empirical findings.

Questions for authors
  1. In line 196, the Bernoulli parameter associated with Bi,tB_{i,t} is denoted by ρi\rho_i. However, unlike πf\pi_f, which is regularized through a KL-divergence term involving rϕ(yt)r_\phi(y_t), there does not appear to be an explicit loss term that regularizes ρi\rho_i. Please clarify why such a regularization term is not required for learning ρi\rho_i.

  2. Appendix E.3 specifies the search spaces for several hyperparameters. I have two related questions.

  • First, does k∈[4,10]k \in [4,10] mean that all seven integer values from 4 to 10 are considered as candidate values?
  • Second, is the same hyperparameter optimization budget applied to the baseline methods? This is an important concern because Appendix H.1 (around lines 1121 onward) appears to indicate that fixed configurations are used for the baselines. If the proposed method is extensively tuned while the baselines use fixed configurations, this could result in an unfair comparison. Please clarify the tuning protocol and optimization budget used for all methods.
5 comments

We thank the reviewer for the careful reading and for recognizing the tractability of the training objective. We appreciate the suggestions on quantitative evaluation, parameter robustness, and the role of the theoretical result.

1. Why private jumps have no separate KL term

The common Bernoulli KL compares the variational common-state posterior with its prior. Private states are instead integrated out in the likelihood:

=(1-\rho_i)t_{\nu_{i,0}}(y_i;f_{\theta i}(z),s_i) +\rho_i t_{\nu_{i,1}}(y_i;f_{\theta i}(z),a_{i,1}s_i).$$ Here, the last argument of each univariate Student-$t$ density is its scale. The ELBO includes the log of this mixture likelihood, and no separate variational approximation to the private state is introduced. As such, there is no additional private-state KL term. In the original experiments, $\rho_i$ is a selected hyperparameter; if learned, it can be optimized through this marginal likelihood. **2. Quantitative synthetic evaluation** Following your suggestion, we computed the two-sample KS distance between standardized return-norm CDFs and report quantitative results across five independently generated S1 datasets. E-VAE-CP has a mean norm KS distance of 0.1230, compared with 0.2597 for Tail-GAN and 0.3166 for Tail-Diffusion. Both comparisons favor E-VAE-CP on every dataset. **Quantile errors and return-norm KS distance on S1.** | Model | QMAE 2.5% | QMAE 5% | QMAE 7.5% | Norm KS | |---|---:|---:|---:|---:| | G-VAE | 0.0669 ± 0.0094 | 0.0442 ± 0.0056 | 0.0613 ± 0.0077 | 0.1214 ± 0.0092 | | E-VAE (no jumps) | 0.0668 ± 0.0107 | 0.0398 ± 0.0047 | 0.0339 ± 0.0027 | 0.1576 ± 0.0070 | | E-VAE-C | 0.0580 ± 0.0061 | 0.0388 ± 0.0039 | 0.0366 ± 0.0033 | 0.1176 ± 0.0054 | | E-VAE-P | 0.0681 ± 0.0104 | 0.0392 ± 0.0055 | 0.0305 ± 0.0018 | 0.1192 ± 0.0094 | | E-VAE-CP | 0.0568 ± 0.0068 | 0.0388 ± 0.0040 | 0.0365 ± 0.0034 | 0.1230 ± 0.0039 | | Tail-GAN | 0.0549 ± 0.0024 | 0.0394 ± 0.0050 | 0.0358 ± 0.0052 | 0.2597 ± 0.0200 | | Tail-Diffusion | 0.2533 ± 0.0120 | 0.2125 ± 0.0086 | 0.1925 ± 0.0121 | 0.3166 ± 0.0090 | | $t^3$-VAE | 0.3934 ± 0.0157 | 0.3686 ± 0.0198 | 0.3466 ± 0.0211 | 0.2996 ± 0.0108 | | Pareto-VAE | 0.1800 ± 0.0261 | 0.2896 ± 0.0409 | 0.3331 ± 0.0418 | 0.3894 ± 0.0296 | Entries are means ± pointwise 95% Student-t confidence-interval half-widths over five independent S1 datasets (N=60, k=5, T=6,000). Norm KS is the two-sample CDF distance for standardized return norms. QMAE is in return percentage points. All measures are smaller-is-better. **3. Parameter sensitivity and joint learning** In the reported implementation, radial hyperparameters are global settings, while the decoder and asset-specific residual scales are learned during training. We varied the common jump probability and jump-component degrees of freedom one at a time and found that the reported metrics were relatively insensitive within the tested ranges. **Local sensitivity of E-VAE-CP on S1.** | $\pi_f$ | $\nu_{f,1}$ | QMAE 2.5% | FD/N | Energy | |---|---:|---:|---:|---:| | 0.010 | 4.5 | 0.0600 ± 0.0093 | 0.0299 ± 0.0031 | 0.0243 ± 0.0007 | | 0.005 | 4.5 | 0.0626 ± 0.0078 | 0.0311 ± 0.0029 | 0.0244 ± 0.0007 | | 0.020 | 4.5 | 0.0562 ± 0.0080 | 0.0286 ± 0.0041 | 0.0251 ± 0.0012 | | 0.010 | 3 | 0.0607 ± 0.0083 | 0.0290 ± 0.0027 | 0.0251 ± 0.0014 | | 0.010 | 6 | 0.0597 ± 0.0072 | 0.0312 ± 0.0046 | 0.0249 ± 0.0022 | Entries are means ± pointwise 95% confidence-interval half-widths over the five datasets. One parameter is changed from the local reference setting (0.01, 4.5) at a time. This is a separate local-sensitivity experiment. We also evaluated a variant that learns selected radial parameters jointly with the neural networks, while keeping the degrees of freedom fixed. Its overall distributional fit is similar to that of the fixed version. **Fixed and learned E-VAE-CP at the local reference setting.** | Radial specification | QMAE 2.5% | FD/N | Energy | |---|---:|---:|---:| | Fixed | 0.0600 ± 0.0093 | 0.0299 ± 0.0031 | 0.0243 ± 0.0007 | | Learned | 0.0649 ± 0.0079 | 0.0307 ± 0.0029 | 0.0241 ± 0.0008 | Entries are means ± pointwise 95% confidence-interval half-widths over the same five datasets. The common jump-component degrees of freedom remain fixed at 4.5; the common jump probability is set to 0.01 in the fixed variant and initialized at 0.01 in the learned variant.

4. The role of Theorem 2

Theorem 2 aims to identify a restriction of the affine factor model. With finite second moments and a nonzero common tail contribution, the common covariance and common tail angular components both span range⁡(A)\operatorname{range}(A). Thus, common variation in the body and the tail shares the same linear factor space. Our empirical analysis, however, finds that the estimated body and tail subspaces do not coincide. This finding suggests that a single linear factor-loading space may be insufficient to describe both ordinary and extreme return variation, and therefore motivates a nonlinear decoder. The controlled decoder ablation further supports this choice: under the same radial specification, nonlinear decoding reduces 2.5% QMAE from 0.916 to 0.448 in the U.S. and from 1.613 to 0.595 in China.

Thank you for your detailed response and additional experiments. I appreciate the authors' efforts in addressing my concerns.

I have two remaining questions.

1. Hyperparameter tuning

While the additional sensitivity analysis is helpful, the hyperparameter tuning budget for the proposed method remains unclear. Could the authors clarify how many hyperparameter configurations were evaluated during the grid search described in Appendix E.3 and compare the tuning budget with those of the baselines?

2. Quantitative synthetic evaluation

The additional quantitative results on S1 are helpful. Since S2 and S3 involve nonlinear factor-to-return mappings, could the authors also provide quantitative results for these settings?

Reviewer_QHYkOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes an elliptical VAE for generating multivariate financial returns with heavy tails and occasional extreme movements. Building on a nonlinear factor model, it separates common market variation from asset-specific residuals and represents both through Gaussian scale mixtures, with additional Student-t mixture components for common and private jumps. An augmented latent-variable formulation yields a standard evidence lower bound with analytic Gaussian, Gamma, and Bernoulli KL terms, while private jump states are marginalized analytically. Generation requires a single decoder evaluation followed by residual sampling. The evaluation covers three synthetic settings, U.S. and Chinese equity markets, several generative baselines, and ablations of jump components. The reported results suggest a promising balance between tail accuracy and overall distributional fit, while a downstream CVaR-constrained portfolio experiment finds that different jump specifications perform best in the two markets.

Critical strengths
  1. The problem formulation is meaningful and well aligned with the application. Separating common-factor extremes from asset-specific extremes provides a useful structural approach to financial scenario generation.

  2. The paper proposes a clean, tractable construction for separate common and idiosyncratic heavy tails. This formulation potentially applies to any Gaussian-scale-mixture latent variable model with a tractable scale KL.

  3. The empirical study includes relevant baselines and a useful decoder ablation.

Critical weaknesses
  1. The methodological advance over existing heavy-tailed VAEs is not yet sufficiently established. The separation of common and asset-specific extremes is well motivated, and the tractable variational formulation seems appealing. However, the paper acknowledges prior VAEs with Student-t priors or observation models, and would benefit from a clearer explanation of which modeling or inference capabilities the proposed combination adds beyond these existing ingredients. The current ablations do not fully resolve this question, since ordinary components use 20 degrees of freedom, whereas jump components use 5, so introducing jumps changes both the mixture structure and the corresponding factor or residual tail index. An independently tuned, single-component Student-t VAE would help distinguish the benefit of explicitly modeling jump states from that of simply choosing heavier-tailed distributions.

  2. The reported results are encouraging, but additional uncertainty analysis would help assess the strength of the conclusions. For example, the U.S. Sharpe ratios of E-VAE-C, E-VAE-P, and E-VAE-CP are relatively close and Table 2 does not report confidence intervals or variability across repeated training runs. The reported summary statistics alone are therefore insufficient to determine whether these differences are robust to sampling and training variability. In addition, Appendix E.3 includes portfolio performance among the selection considerations, but the selection dates and precise decision rule are not fully specified. Sensitivity analyses would also be useful, particularly because some selected values lie at the search boundaries.

  3. The degrees of freedom, mixture probabilities, and relative jump scales are selected as hyperparameters and held fixed during training, rather than estimated jointly with the network. The decoder could still learn nonlinear dependence and tail-relevant structure, so this is not equivalent to merely rescaling a fixed return distribution. Nevertheless, the paper would benefit from distinguishing the tail characteristics specified through these hyperparameters from those learned through the decoder, and from examining sensitivity to the chosen tail specifications.

  4. The factorized likelihood and single-pass generation are attractive computational properties, and the paper explains that private-state marginalization scales linearly with the number of assets. However, the real-data experiments involve 95 and 158 assets, leaving scalability to substantially larger universes unclear. Experiments varying the number of assets, latent factors, and generated scenarios would help characterize training time, sampling throughput, peak memory usage, and tail-modeling accuracy at larger scales.

Questions for authors
  1. Given that the ordinary and jump components use 20 and 5 degrees of freedom, respectively, what evidence separates the benefit of introducing mixture states from that of introducing heavier-tailed components? Do the improvements persist against a no-jump E-VAE whose factor and residual degrees of freedom are independently tuned, including comparably heavy-tailed specifications, under the same architecture, standard ELBO, and validation protocol?

  2. Could the authors specify the exact dates or windows used for the grid search in Appendix E.3 and the decision rule for combining tail accuracy, distributional fit, and portfolio performance? When was the final, globally fixed configuration selected relative to the first reported test window?

  3. How robust is the conclusion that common jumps are more beneficial in the U.S. and private jumps in China to training and scenario-generation randomness? With the degrees of freedom, mixture probabilities, and relative jump scales held fixed during training, which characteristics of the generated marginal and joint tails are learned through the decoder and residual scales, and which remain constrained by the selected radial distributions?

  4. How does the proposed model depend on the number of assets, latent factors, and generated scenarios?

2 comments

We thank the reviewer for the thoughtful assessment and for recognizing the financial motivation and tractable inference of E-VAE. We appreciate the suggestions on mixture modeling, learned tail behavior, and parameter robustness. The additional results below address these points.

1. Common-factor mixture ablation

Our model allows the common and private distributions to have separate mixture probabilities, relative scales, and component tails. Under the proposed framework, Gaussian scale-mixture augmentation gives a standard ELBO with analytic KL terms for the common latent variables, while private mixture states are marginalized analytically.

To examine the role of mixtures beyond changing degrees of freedom, we compared four common-factor specifications: A uses a single Student-tt component; B mixes scales with both component degrees of freedom fixed at 8; C mixes degrees of freedom at equal scatter; and D allows both scale and degrees of freedom to differ. All four retain the same private radial specification, with each model fitted separately.

Common-factor ablation: tail-quantile errors.

SpecificationQMAE 2.5%QMAE 5%QMAE 7.5%
A: single component0.0641 ± 0.00760.0527 ± 0.00570.0500 ± 0.0084
B: scale mixture0.0584 ± 0.00970.0372 ± 0.00430.0331 ± 0.0028
C: df mixture0.0654 ± 0.00650.0383 ± 0.00420.0317 ± 0.0016
D: scale + df mixture0.0568 ± 0.00680.0388 ± 0.00400.0365 ± 0.0034

Common-factor ablation: overall distributional fit.

SpecificationFD/NEnergySW1Norm KS
A: single component0.0285 ± 0.00360.0261 ± 0.00220.0653 ± 0.00200.1215 ± 0.0109
B: scale mixture0.0298 ± 0.00260.0238 ± 0.00140.0615 ± 0.00210.1187 ± 0.0078
C: df mixture0.0324 ± 0.00450.0248 ± 0.00210.0630 ± 0.00250.1216 ± 0.0077
D: scale + df mixture0.0302 ± 0.00430.0257 ± 0.00270.0640 ± 0.00260.1230 ± 0.0039

Both tables report means ± pointwise 95% Student-t confidence-interval half-widths over five independent S1 datasets (N=60, k=5, T=6,000). All measures are smaller-is-better. Compared with A, B improves 5% and 7.5% QMAE on all five datasets.

2. Specified parameters, sensitivity, and joint learning

In the fixed specification, radial hyperparameters determine the component distribution families, mixture probabilities, and relative jump scales. The nonlinear factor-to-return map and asset-specific residual scales are learned from data. They determine how factor variation is transmitted across assets and the magnitudes of residual fluctuations within the selected radial families. Following your helpful suggestion, we also varied the common jump probability and jump-component degrees of freedom one at a time. The reported metrics remain relatively insensitive within the tested ranges.

Local sensitivity of E-VAE-CP on S1.

πf\pi_fνf,1\nu_{f,1}QMAE 2.5%FD/NEnergy
0.0104.50.0600 ± 0.00930.0299 ± 0.00310.0243 ± 0.0007
0.0054.50.0626 ± 0.00780.0311 ± 0.00290.0244 ± 0.0007
0.0204.50.0562 ± 0.00800.0286 ± 0.00410.0251 ± 0.0012
0.01030.0607 ± 0.00830.0290 ± 0.00270.0251 ± 0.0014
0.01060.0597 ± 0.00720.0312 ± 0.00460.0249 ± 0.0022

Entries are means ± pointwise 95% confidence-interval half-widths over the five datasets. One parameter is changed from the local reference setting (0.01, 4.5) at a time.

We also evaluated a variant that learns selected radial parameters jointly with the neural networks, while keeping the degrees of freedom fixed. Its overall distributional fit is similar to that of the fixed version.

Fixed and learned E-VAE-CP at the local reference setting.

Radial specificationQMAE 2.5%FD/NEnergy
Fixed0.0600 ± 0.00930.0299 ± 0.00310.0243 ± 0.0007
Learned0.0649 ± 0.00790.0307 ± 0.00290.0241 ± 0.0008

Entries are means ± pointwise 95% confidence-interval half-widths over the same five datasets. The common jump-component degrees of freedom remain fixed at 4.5; the common jump probability is set to 0.01 in the fixed variant and initialized at 0.01 in the learned variant.

I thank the authors for the additional experiments. The concerns on the contribution over existing heavy-tailed VAEs and the uncertainty analysis are mostly addressed. However, the issues on hyperparameter selection remain unclear. A configuration might remain fixed throughout the reported evaluation but still have been selected using validation data from dates later than the first test window. Keeping the configuration fixed afterward does not eliminate this chronological issue. In addition, the concerns regarding scalability remain unresolved. Therefore, I will maintain my current score.

Reviewer_hNagOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes an Elliptical Variational Auto-Encoder (E-VAE) for generating heavy-tailed financial returns. It develops a nonlinear elliptical factor model that captures heavy-tailed common factors and idiosyncratic risks, with separate mechanisms for common and asset-specific jumps. The paper also theoretically characterizes the limitations of Gaussian VAEs with globally Lipschitz decoders in reproducing power-law portfolio tails. The proposed model is evaluated on synthetic data and U.S. and Chinese stock returns against several generative baselines, including Gaussian VAE, t³-VAE, Pareto-VAE, Tail-GAN, and Tail-Diffusion. E-VAE variants consistently achieve better overall distributional fit than Tail-GAN and Tail-Diffusion in both markets, while tail-quantile rankings vary across markets and quantiles. A downstream CVaR-constrained portfolio experiment further evaluates the practical value of the generated scenarios, where E-VAE-C achieves the best Sharpe ratio and cumulative net excess return in the U.S. and E-VAE-P performs best in China, suggesting that common and asset-specific jumps have different portfolio implications across the two markets.

Critical strengths

1: Theorem 1 establishes that under fixed parameters and a globally Lipschitz decoder, a Gaussian VAE generates sub-Gaussian centered portfolio returns and therefore cannot reproduce power-law portfolio tails. The proof provides a mathematical rationale for the proposed heavy-tailed modeling approach, rather than relying solely on empirical motivation. Although the result is conditional on its assumptions, it is the paper's clearest theoretical contribution.

2: The model distinguishes common jumps from asset-specific jumps, corresponding to economically meaningful sources of extreme returns. The paper evaluates these components not only through distributional metrics on synthetic and U.S. and Chinese stock-return data, but also in a CVaR-constrained portfolio task. The reported results suggest that the relative usefulness of common and private jump components differs between the two markets under the studied experimental setup. This provides evidence that the proposed decomposition may have value beyond fitting return distributions, strengthening the paper's practical relevance.

Critical weaknesses

1: The paper’s headline claims are deliberately narrow: E-VAE variants achieve better overall distributional fit (FD/N, Energy, SW1) than Tail-GAN and Tail-Diffusion in both markets, while the cross-market portfolio results are presented as suggestive. Even within this restricted claim, however, the evidence is uneven: in the Chinese market, t³-VAE achieves lower FD/N, Energy, and SW1 values than the E-VAE variants, and G-VAE achieves the lowest QMAE at the 7.5% quantile in both markets. The reported metrics are presented as point estimates, without confidence intervals, standard errors, or formal statistical comparisons establishing that the differences between methods are meaningful rather than noise. The portfolio evaluation additionally relies on a single CVaR-constrained allocation setup, so the robustness of the reported portfolio implications across alternative settings is untested.

2: The theoretical results do not fully characterize the deployed nonlinear model, and the robustness of the selected hyperparameters remains uncertain. The theoretical results on common covariance and extreme-return directions rely on an affine factor-to-return specification, whereas the proposed E-VAE uses a nonlinear neural decoder. Although the paper provides a decoder ablation showing empirical improvements from nonlinear decoding, this does not establish that the theoretical characterizations extend to the nonlinear model. The paper also conducts a restricted hyperparameter grid search, but the stability of its conclusions beyond the selected configurations remains unclear. The gap between the formal results and the deployed architecture, together with the remaining uncertainty about hyperparameter robustness, limits my confidence in the scope and generality of the contribution.

Questions for authors

1: Appendix H.1 states that t³-VAE and Pareto-VAE are used through “tabular adaptations of the publicly released implementations”. Since the E-VAE variants’ gains are partly reported relative to these baselines, could the authors clarify whether the baselines’ hyperparameters were tuned on the validation sets under the same protocol used for E-VAE, and how the applied corrections affect the comparison?

2: Could the authors provide appropriate uncertainty quantification or formal statistical comparisons for the main empirical claims, particularly comparisons between E-VAE variants and competing methods on tail-quantile and distributional metrics, as well as the leading Sharpe ratios? Since the reported performance varies across markets and metrics, such evidence would help establish how reliably the observed differences support the paper’s empirical conclusions.

  1. Section E.3 states that preserving the latent tail index additionally requires a non-degenerate asymptotic decoder response along directions carrying positive tail mass. Is this condition supported by a formal result for the nonlinear decoder used in the experiments, or is it intended as a heuristic or design motivation?
1 comment

We thank the reviewer for the careful assessment and for recognizing both the theoretical motivation for heavy-tailed generation and the financial relevance of separating common and private extremes. We appreciate the suggestions on repeated evaluation, parameter sensitivity, and the connection between the theory and the nonlinear decoder.

1. Repeated evaluation and uncertainty

Following your suggestion, we repeated S1 on five independently generated datasets, evaluating the full 64-window horizon for each. The general response reports model-wise confidence intervals; the table below gives paired differences on the same datasets. E-VAE-CP has tail-quantile errors close to Tail-GAN's and smaller FD/N, Energy, SW1, and norm KS distances on all five datasets. The paired intervals are able to quantify these improvements.

Paired differences: E-VAE-CP minus the comparator.

Measurevs. Tail-GANvs. Tail-Diffusion
QMAE 2.5%0.0019 [-0.0053, 0.0090]-0.1965 [-0.2126, -0.1804]
QMAE 5%-0.0006 [-0.0074, 0.0061]-0.1737 [-0.1835, -0.1639]
QMAE 7.5%0.0007 [-0.0028, 0.0042]-0.1560 [-0.1649, -0.1472]
FD/N-0.2182 [-0.2349, -0.2014]-0.3059 [-0.3148, -0.2971]
Energy-0.0547 [-0.0618, -0.0476]-0.0795 [-0.0828, -0.0761]
SW1-0.0529 [-0.0586, -0.0472]-0.0800 [-0.0829, -0.0771]
Norm KS-0.1367 [-0.1543, -0.1192]-0.1936 [-0.2021, -0.1851]

Entries are mean differences with pointwise 95% Student-t confidence intervals in brackets, paired across five independent S1 datasets (N=60, k=5, T=6,000). Negative values favor E-VAE-CP.

2. Why the affine theorem motivates nonlinear decoding

Indeed, the role of Theorem 2 is to identify a restriction imposed by a linear decoder and help to empirically support a non-linear decoder. With finite second moments and a nonzero common tail contribution, the common covariance and common tail angular components both span range⁡(A)\operatorname{range}(A). Ordinary and extreme common variation therefore share a fixed linear space.

In our empirical analysis, the nonlinear decoder relaxes this restriction. The body–tail geometry motivates this flexibility, and the controlled decoder ablation measures its benefit. Under the same radial specification, nonlinear decoding reduces 2.5% QMAE from 0.916 to 0.448 in the U.S. and from 1.613 to 0.595 in China. This ablation provides direct empirical support for nonlinear decoding.

3. Parameter sensitivity and joint learning

Thank you for suggesting this sensitivity analysis. We varied the common jump probability and jump-component degrees of freedom one at a time and found that the reported metrics were relatively insensitive within the tested ranges.

Local sensitivity of E-VAE-CP on S1.

πf\pi_fνf,1\nu_{f,1}QMAE 2.5%FD/NEnergy
0.0104.50.0600 ± 0.00930.0299 ± 0.00310.0243 ± 0.0007
0.0054.50.0626 ± 0.00780.0311 ± 0.00290.0244 ± 0.0007
0.0204.50.0562 ± 0.00800.0286 ± 0.00410.0251 ± 0.0012
0.01030.0607 ± 0.00830.0290 ± 0.00270.0251 ± 0.0014
0.01060.0597 ± 0.00720.0312 ± 0.00460.0249 ± 0.0022

Entries are means ± pointwise 95% confidence-interval half-widths over the five datasets. Only one parameter is changed from the local reference setting (0.01, 4.5) at a time.

We also evaluated a variant that learns selected radial parameters jointly with the neural networks, while keeping the degrees of freedom fixed. Its overall distributional fit is similar to that of the fixed version, as shown below.

Fixed and learned E-VAE-CP at the local reference setting.

Radial specificationQMAE 2.5%FD/NEnergy
Fixed0.0600 ± 0.00930.0299 ± 0.00310.0243 ± 0.0007
Learned0.0649 ± 0.00790.0307 ± 0.00290.0241 ± 0.0008

Entries are means ± pointwise 95% confidence-interval half-widths over the same five independent datasets. The common jump-component degrees of freedom remain fixed at 4.5 in both variants.

Comments on the paper

1 comment

We thank all reviewers for their careful reading and constructive feedback. We appreciate their recognition of the financial motivation and the tractable separation of common and asset-specific extremes. Their comments have helped us strengthen the experimental evidence and clarify the modeling contribution.

In response, we repeated the S1 comparison on five independently generated datasets and report confidence intervals below. Each repetition covers the full 64-window evaluation. E-VAE-CP retains tail-quantile accuracy close to Tail-GAN while providing a substantially better overall distributional fit. Its FD/N, Energy, SW1, and return-norm KS distances are smaller than Tail-GAN's on all five datasets. The comparison with Tail-Diffusion also favors E-VAE-CP across all reported measures.

We additionally examined sensitivity to the common jump probability and degrees of freedom, and compared single-component and mixture distributions for the common factors. The ablation shows that a scale mixture can improve tail accuracy even when both components have the same degrees of freedom. Overall fit also remains stable under the tested local parameter changes. We provide these results in the individual responses.

Table 1. Tail-quantile errors on five independent S1 datasets.

ModelQMAE 2.5%QMAE 5%QMAE 7.5%
G-VAE0.0669 ± 0.00940.0442 ± 0.00560.0613 ± 0.0077
E-VAE (no jumps)0.0668 ± 0.01070.0398 ± 0.00470.0339 ± 0.0027
E-VAE-C0.0580 ± 0.00610.0388 ± 0.00390.0366 ± 0.0033
E-VAE-P0.0681 ± 0.01040.0392 ± 0.00550.0305 ± 0.0018
E-VAE-CP0.0568 ± 0.00680.0388 ± 0.00400.0365 ± 0.0034
Tail-GAN0.0549 ± 0.00240.0394 ± 0.00500.0358 ± 0.0052
Tail-Diffusion0.2533 ± 0.01200.2125 ± 0.00860.1925 ± 0.0121
t3t^3-VAE0.3934 ± 0.01570.3686 ± 0.01980.3466 ± 0.0211
Pareto-VAE0.1800 ± 0.02610.2896 ± 0.04090.3331 ± 0.0418

Table 2. Overall distributional fit on the same five datasets.

ModelFD/NEnergySW1Norm KS
G-VAE0.0409 ± 0.00390.0265 ± 0.00180.0653 ± 0.00310.1214 ± 0.0092
E-VAE (no jumps)0.0346 ± 0.00240.0300 ± 0.00360.0702 ± 0.00290.1576 ± 0.0070
E-VAE-C0.0287 ± 0.00440.0251 ± 0.00180.0628 ± 0.00250.1176 ± 0.0054
E-VAE-P0.0322 ± 0.00400.0256 ± 0.00310.0643 ± 0.00300.1192 ± 0.0094
E-VAE-CP0.0302 ± 0.00430.0257 ± 0.00270.0640 ± 0.00260.1230 ± 0.0039
Tail-GAN0.2484 ± 0.01740.0804 ± 0.00730.1170 ± 0.00660.2597 ± 0.0200
Tail-Diffusion0.3361 ± 0.00940.1052 ± 0.00190.1440 ± 0.00150.3166 ± 0.0090
t3t^3-VAE0.0411 ± 0.00180.0727 ± 0.00470.1228 ± 0.00230.2996 ± 0.0108
Pareto-VAE0.0663 ± 0.01390.1285 ± 0.02170.1413 ± 0.01190.3894 ± 0.0296

Entries are means ± pointwise 95% confidence-interval half-widths over five independent S1 datasets, calculated using Student-t intervals with four degrees of freedom. QMAE is in return percentage points. All measures are smaller-is-better.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.